Pattern: VDSim as the ego plant inside an external benchmark

Replacing a kinematic ego model with a friction-aware plant, and changing nothing else.

This is a usage pattern. It makes no benchmark claim and shows no result numbers: the repository does not yet contain a script that reproduces one, and a number without a reproduce command is not a result (Validation).

The setup

Many closed-loop driving benchmarks propagate the ego vehicle with a kinematic bicycle model. One example is NeuroNCAP¹, an open closed-loop safety benchmark whose authors state that its ego model ignores delays, friction, suspension and road surface. That is a reasonable simplification for a perception benchmark. It is also exactly the assumption VDSim exists to test: a manoeuvre the kinematic ego executes perfectly may be infeasible on real tyres at real grip.

The idea

Keep the whole loop — renderer, policy, controller, scenarios — and swap only the block that turns a command into motion.

flowchart LR
  A["perception"] --> B["policy"] --> C["controller"]
  C -- "steer, accel" --> D["ego dynamics<br/>kinematic bicycle → VDSim plant"]
  D -- "next ego pose" --> E["render next frame"]
  E --> A

Everything else is unchanged, so any difference in outcome is attributable to the dynamics substitution alone.

What to measure, in order

Keep the claims in separate layers and do not quote a higher layer from a lower one.

Layer	Question	What it needs
1 — open-loop replay	Is an identical steer profile feasible on the kinematic model but not on VDSim at lower μ?	one recorded steer profile, a μ sweep
2 — simple closed loop	Does a feedback planner driving VDSim hit a μ-dependent feasibility boundary?	a stub planner in the loop
3 — full benchmark	Does a real driving policy's benchmark score change under VDSim dynamics?	the full rendered closed loop

The expected direction is that lower grip shrinks the avoidance margin the kinematic ego reports. The size of that effect depends on vehicle parameters and on the manoeuvre, so publish it only together with the configuration and a reproduce command.

---

1. A. Tonderski et al., NeuroNCAP: Photorealistic Closed-loop Safety Testing for Autonomous Driving, arXiv:2404.07762. NeuroNCAP is MIT-licensed. This page is an independent usage note and is not affiliated with or endorsed by its authors. ↩